ntropi

Python SDK and runtime for the Entropi stack. One import to run and operate the whole stack from Python — a full dev tool to develop, teach, trace, and connect to the nbot ecosystem.

It wraps the full surface — the HTTP API, every slash/utility command, config, and process lifecycle — with typed models, sync and async clients, and a runtime that can launch the services (or connect to running ones).

Quickstart

import ntropi

# develop: define a (sovereign) node's pool + config from Python — no TOML editing
cfg = ntropi.nbot.Config.create("/my/project", node_id="rtx")
cfg.pool.add("gemma", endpoint="https://api.infercom.ai/v1",
             model="gemma-4-31B-it", api_key="env:INFERCOM_APIKEY",
             modalities=["image_in"])
cfg.set_network(encrypt=True).set_rl("thompson")
cfg.set_security(allowed_paths=["/data"])          # jurisdiction / data-in-place
cfg.save()

# connect to a running node (auto-discovers .nconfig/control.json)
c = ntropi.nbot.Client(project="/my/project")

for m in c.pool.list():
    print(m.id, m.capacity, m.modalities)

# trace: run and observe a turn as typed events
for ev in c.tasks.stream("Plot the data in report.csv", new_session=True):
    if ev.is_token:
        print(ev.text, end="")

# teach: freeze learning for a clean eval, inspect the policy, give feedback
c.rl.set_learning(False)
print(c.routing.policies()["active"])
c.rl.feedback("s-123", "positive")

# connect: blueprints, connectors, mesh — all typed
c.blueprints.run("kyc-check", {"customer_id": "42"})
print(c.connectors.list())

# every slash-command is a typed method (a bare string is a message)
print(c.command.pool())
print(c.command.rl())

# drop-in OpenAI-compatible facade
c.chat.completions.create(model="nbot", messages=[{"role": "user", "content": "hi"}])

Everything above has an async twin with an identical surface — for all three components:

async with ntropi.nbot.AsyncClient(project="/my/project") as ac:
    await ac.tasks.submit("hello")
    print(await ac.rl.learning())

async with ntropi.horizon.AsyncClient("http://127.0.0.1:8848") as ah:
    job = await ah.rl.env_train(prompt="do X", iterations=5)
    async for point in job.stream():
        print(point["reward"])

Launch the stack from Python

# start nbot (+ Horizon UI, + the NBotWeb chat gateway) for a project
stack = ntropi.launch("/my/project", with_web=True)
print(stack.nbot.health())
# … Horizon is live on :8848, the web gateway on :8080 …
stack.close()   # stops only what it started

Build a chat app on the gateway

w = ntropi.web.Client("http://127.0.0.1:8080")
w.tasks.submit("summarise the attached report", new_session=True)
key = w.admin.keys.create("mobile-app")     # secret returned once
w.admin.users.create("dev@team.com", role="user", projects=["/my/project"])

From the shell (pip install 'ntropi[cli]')

ntropi launch ./my-project --web
ntropi pool --project ./my-project
ntropi chat "summarise report.csv" --project ./my-project
ntropi rl-learning off --project ./my-project      # freeze for eval

Install

pip install ntropi

Extras: "ntropi[cli]" adds the ntropi command. The gateway extra pulls in nbot-gateway, which is not published yet, so leave it off until it is.

The same wheels are mirrored on the platform's own index for air-gapped installs: pip install --extra-index-url https://platform.entropi.ai/pip/ ntropi.

Install (dev)

python -m venv .venv && . .venv/bin/activate
pip install -e ".[dev,cli]"
pytest -q

Requires Python ≥ 3.10. ntropi drives the nbot / nbot-horizon / nbot-web binaries — it does not bundle them; install them and put them on PATH (or set NTROPI_NBOT_BIN / NTROPI_HORIZON_BIN / NTROPI_WEB_BIN).

Status

0.5.0 — full coverage across the stack:

• nbot — pool, projects, sessions, tasks, chat (+ OpenAI facade), models, the typed command namespace, and the structured surface (node RL, blueprints, kb, research, artifacts, prompts, notes, connectors, routing, agents, satellites, trace, files, manifesto).
• Horizon — suites, datasets, RL train/env-train (Job), sessions/traces.
• NBotWeb — the full gateway: auth, tasks/chat, read surfaces, adapters, admin (keys/users/settings), MCP, proxy.
• Sync and async everywhere; typed Config; runtime launch; a ntropi CLI.

See CHANGELOG.md and docs/plans.